怎麼有效率地找到適合合作的 Instagram 創作者

#!/usr/bin/env python3
"""Find candidate IG handles via Startpage + Yahoo Taiwan (condensed from scripts/search_harvest.py).
Usage: python3 ig_search.py queries.tsv out.json   # queries.tsv: sub_topic<TAB>query per line
Rules: 4-8 s between searches; an engine that shows a captcha is dropped for the rest of the run.
"""
import json, random, re, sys, time, urllib.parse
from playwright.sync_api import sync_playwright

ENGINES = {"startpage": "https://www.startpage.com/do/search?q={}&language=zh-TW",
           "yahoo": "https://tw.search.yahoo.com/search?p={}&n=40"}
SKIP = {"p", "reel", "reels", "explore", "accounts", "stories", "tv", "popular", "about", "legal", "direct", "web", "developer"}
H = r"[A-Za-z0-9._]{2,30}"
BLOCK = re.compile(r"captcha|unusual traffic|are you a robot|異常流量|請證明您不是機器人", re.I)

queries = [l.rstrip("\n").split("\t", 1) for l in open(sys.argv[1], encoding="utf-8") if "\t" in l]
results, dead = [], set()
with sync_playwright() as p:
    b = p.chromium.launch(headless=True)
    ctx = b.new_context(locale="zh-TW")       # no cookies, not logged in
    for topic, q in queries:
        for eng, tmpl in ENGINES.items():
            if eng in dead:
                continue
            pg = ctx.new_page()
            pg.goto(tmpl.format(urllib.parse.quote(q)), wait_until="domcontentloaded", timeout=30000)
            pg.wait_for_timeout(2500)
            if BLOCK.search(pg.content()) and "instagram.com" not in pg.content():
                dead.add(eng); pg.close(); continue
            links = pg.evaluate("""() => [...document.querySelectorAll('a[href]')].map(a => {
                let h = a.getAttribute('href') || ''; const m = h.match(/RU=([^/]+)/);
                if (m) { try { h = decodeURIComponent(m[1]); } catch (e) {} }
                const box = a.closest('li, .w-gl__result, .result, div.dd, div.algo') || a.parentElement;
                return {href: h, text: (box ? box.innerText : '').slice(0, 600)};
            }).filter(x => x.href.includes('instagram.com'))""")
            for it in links:
                found = set(re.findall(r"\(@(" + H + r")\)", it["text"])) | set(re.findall(r"(?<![\w.])@(" + H + r")", it["text"]))
                m = re.search(r"instagram\.com/(" + H + r")", it["href"])
                if m:
                    found.add(m.group(1))
                post = re.search(r"instagram\.com/(?:" + H + r"/)?(?:p|reel)/([\w-]+)", it["href"])
                handles = sorted({f.lower().strip(".") for f in found} - SKIP)
                results.append({"topic": topic, "engine": eng, "query": q, "url": it["href"],
                                "handles": handles, "post": post.group(1) if post else None})
            pg.close()
            time.sleep(random.uniform(4, 8))
    b.close()
json.dump(results, open(sys.argv[2], "w", encoding="utf-8"), ensure_ascii=False, indent=1)
print("dead engines:", dead)

#!/usr/bin/env python3
"""Logged-out public IG profile check (condensed from scripts/ig_public.py).
Usage: pip install playwright && playwright install chromium
       python3 ig_verify.py OUTDIR handle1 handle2 ...
Rules: 3-8 s between pages, max 150 pages per round, stop on login wall / 429.
"""
import datetime, json, random, re, sys, time
from pathlib import Path
from playwright.sync_api import sync_playwright

B64 = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-_"
EMAIL = re.compile(r"[\w.+-]+@[\w-]+\.[\w.]+")
PHONE = re.compile(r"(?<!\d)(?:\+?886[-\s]?|0)9\d{2}[-\s]?\d{3}[-\s]?\d{3}(?!\d)")
redact = lambda s: PHONE.sub("[phone]", EMAIL.sub("[email]", s or ""))

def shortcode_date(sc):                      # post shortcode -> post date, no extra page load
    n = 0
    for c in sc[:11]:
        n = n * 64 + B64.index(c)
    return datetime.date.fromtimestamp(((n >> 23) + 1314220021721) / 1000).isoformat()

out = Path(sys.argv[1]); out.mkdir(parents=True, exist_ok=True)
handles = list(dict.fromkeys(h.lower().lstrip("@") for h in sys.argv[2:]))[:150]
res, stop = [], None
with sync_playwright() as p:
    b = p.chromium.launch(headless=True)
    ctx = b.new_context(locale="zh-TW")       # fresh context: no cookies, not logged in
    for i, h in enumerate(handles):
        if i:
            time.sleep(random.uniform(3, 8))
        pg = ctx.new_page(); r = {"handle": h}
        resp = pg.goto(f"https://www.instagram.com/{h}/", wait_until="domcontentloaded", timeout=45000)
        pg.wait_for_timeout(3500)
        body = pg.evaluate("() => document.body ? document.body.innerText.slice(0, 1500) : ''")
        r["followers"] = pg.evaluate("() => (document.querySelector('meta[property=\"og:description\"]')||{}).content || null")
        r["bio"] = redact(body)
        codes = pg.evaluate("() => [...document.querySelectorAll('a[href*=\"/p/\"],a[href*=\"/reel/\"]')]"
                            ".map(a => (a.getAttribute('href').match(/\\/(?:p|reel)\\/([\\w-]+)/)||[])[1]).filter(Boolean)")
        r["latest_post"] = max((shortcode_date(c) for c in codes[:12]), default=None)
        pg.screenshot(path=str(out / f"{h}.png"))
        if (resp and resp.status == 429) or "accounts/login" in pg.url or "請稍候幾分鐘" in body:
            stop = f"blocked at {h}"; r["blocked"] = True   # stop the run; no retry, no IP change
        elif "無法顯示" in body or "此連結可能故障" in body:
            r["not_found"] = True
        res.append(r); pg.close()
        if stop:
            break
    b.close()
(out / "ig_public.json").write_text(json.dumps(res, ensure_ascii=False, indent=1))
print(f"pages={len(res)} stop={stop}")

Goal: collect as many CANDIDATE Instagram creators as possible who may match:
- based in [COUNTRY, e.g. Taiwan], posting in [LANGUAGE, e.g. Traditional Chinese]
- about [MIN, e.g. 1,000]–[MAX, e.g. 50,000] followers
- topic: [TOPIC, e.g. 讀書帳 / studygram / 學習方法 / 筆記 / 自律]

Do NOT filter or verify follower counts. I will verify every account myself. Include every handle you find, even if uncertain.
Use public sources only; never log in, never create accounts, never follow/like/comment/message anyone. Do not ask me questions.

Split the search into these sub-topics and cover each one (in parallel sub-agents / a parallel workflow if you can):
 1. [SUB-TOPIC 1, e.g. 學測 / 分科測驗 / 高中讀書帳]
 2. [SUB-TOPIC 2, e.g. 會考 / 國中讀書帳]
 3. [SUB-TOPIC 3, e.g. 國考 / 研究所 / 醫學系・藥學系讀書]
 4. [SUB-TOPIC 4, e.g. 多益 / 雅思 / 語言學習帳]
 5. [SUB-TOPIC 5, e.g. 筆記 / 手帳 / 文具 / iPad 筆記]
 6. [SUB-TOPIC 6, e.g. 大學生讀書日常 / 自律 / study with me]
For each sub-topic use many query variants, e.g. `site:instagram.com [KEYWORD] <sub-topic>`, `site:threads.com [KEYWORD] <sub-topic>`, plus Dcard / PTT / blog roundups, public KOL stats sites (e.g. Kolr, StarNgage), and links between accounts (campus ambassador lists, shout-outs, "friends" highlights). You may write code in your sandbox to extract handles in bulk.

Exclude these already-known accounts: [LIST, or "none"].

Deliverable (attach the file): candidates.csv with columns exactly:
ig_username, sub_topic, source_url, reported_follower_count (as seen in the source, may be wrong), country_hint, topic_hint
Also give a short summary: queries per sub-topic and number of unique candidates.

{
  "workflow": "instagram-kol-search-by-criteria",
  "version": "2026-10-17-v3",
  "context": "MAR-903 Forest GTM. Example criteria: Taiwan, Traditional Chinese, study topic (讀書帳 / studygram), 1,000-50,000 followers, latest post within 12 months.",
  "decision": {
    "default": "method_1_search_engine_plus_local_verification (no Manus needed)",
    "optional_add_on": "manus_article_candidates: run once only to pick up candidates from roundup articles, interviews and third-party KOL sites; send its handles to method_1 step_2 verification",
    "scale_needed_and_budget_approved": "method_3_database_tools",
    "creator_market_in_creator_marketplace_list": "method_2_instagram_creator_marketplace",
    "taiwan_in_creator_marketplace": "not in the 18-market list (as of 2026-03); a human must check Meta Business Suite"
  },
  "guardrails": [
    "Never log in to Instagram or Meta, never create accounts, never follow, like, comment or message anyone.",
    "Never print, log or commit API keys or tokens. Manus key is read only from env MANUS_API_KEY.",
    "Never write creator emails or phone numbers into any output; redact them as [email] / [phone].",
    "Instagram login wall (redirect to accounts/login), the message 請稍候幾分鐘, or HTTP 429: stop the whole run, record where it stopped, do not retry, do not change IP, do not bypass.",
    "Search engine shows a captcha: drop that engine for the rest of the run.",
    "Manus: never buy credits or upgrade, never auto-accept high-credit confirmations."
  ],
  "method_1_search_engine_plus_local_verification": {
    "runtime": "Machine on a home or office network (tested on a Mac mini). Python 3 + Playwright Chromium, headless, fresh browser context with no cookies, not logged in, locale zh-TW.",
    "step_1_discover": {
      "engines": {
        "startpage": "https://www.startpage.com/do/search?q={query}&language=zh-TW",
        "yahoo_tw": "https://tw.search.yahoo.com/search?p={query}&n=40"
      },
      "engine_notes": "Startpage returns Google results and usually links profile pages; Yahoo Taiwan returns Bing results and mostly links post pages. google.com, Brave and DuckDuckGo show captchas or error pages in headless mode.",
      "do_not_use": "Claude Code built-in WebSearch for Instagram discovery: it ignores site:instagram.com (tested: 144 queries, 14 candidates, 1 pass).",
      "interval_seconds": [
        4,
        8
      ],
      "query_pattern": "site:instagram.com 讀書帳 <keyword>",
      "sub_topics": {
        "t1_gsat": [
          "site:instagram.com 讀書帳 學測",
          "site:instagram.com 讀書帳 GSAT",
          "site:instagram.com 讀書帳 116學測",
          "site:instagram.com 讀書帳 117學測",
          "site:instagram.com 讀書帳 分科測驗",
          "site:instagram.com 讀書帳 高中生 筆記"
        ],
        "t2_cap": [
          "site:instagram.com 讀書帳 會考",
          "site:instagram.com 讀書帳 CAP 國中",
          "site:instagram.com 讀書帳 117會考",
          "site:instagram.com 讀書帳 118會考",
          "site:instagram.com 國中 讀書帳 筆記分享",
          "site:threads.com 讀書帳 會考"
        ],
        "t3_exam": [
          "site:instagram.com 讀書帳 國考",
          "site:instagram.com 讀書帳 醫學系",
          "site:instagram.com 讀書帳 藥學系",
          "site:instagram.com 讀書帳 研究所",
          "site:instagram.com 讀書帳 公職 高普考",
          "site:instagram.com 讀書帳 律師 國考"
        ],
        "t4_lang": [
          "site:instagram.com 讀書帳 多益",
          "site:instagram.com 讀書帳 雅思",
          "site:instagram.com 讀書帳 留學",
          "site:instagram.com 讀書帳 日檢",
          "site:instagram.com 英文學習 台灣 帳號 筆記",
          "site:instagram.com 讀書帳 英文 單字"
        ],
        "t5_notes": [
          "site:instagram.com 讀書帳 筆記 iPad",
          "site:instagram.com 讀書帳 goodnotes",
          "site:instagram.com 讀書帳 手帳",
          "site:instagram.com 讀書帳 文具",
          "site:instagram.com 讀書帳 筆記分享",
          "site:instagram.com 讀書帳 collanote"
        ],
        "t6_uni": [
          "site:instagram.com 讀書帳 大學生",
          "site:instagram.com 讀書帳 台大",
          "site:instagram.com 讀書帳 自律",
          "site:instagram.com 讀書帳 讀書縮時",
          "site:instagram.com studygram 台灣 讀書帳",
          "site:instagram.com 讀書帳 成大單車節校園大使"
        ]
      },
      "volume_tested": "36 queries x 2 engines = 72 searches, 11 minutes, no captcha",
      "extract": [
        "handle from instagram.com/<handle>/ or threads.com/@<handle> URLs",
        "handles in titles/snippets: '(@handle)', '<handle> on Instagram', '@handle'",
        "post shortcodes from instagram.com/p/<code>/ and /reel/<code>/ URLs"
      ],
      "skip_path_segments": [
        "p",
        "reel",
        "reels",
        "explore",
        "accounts",
        "stories",
        "tv",
        "popular",
        "about",
        "legal",
        "direct",
        "web",
        "developer"
      ],
      "resolve_posts": "Open https://www.instagram.com/p/<code>/ logged out and read the owner handle from the page meta tags. Same 3-8 s interval; these pages count toward the 150-page cap.",
      "normalize": "lowercase, strip @, de-duplicate, remove handles already verified in earlier rounds"
    },
    "step_2_verify": {
      "url": "https://www.instagram.com/<handle>/",
      "interval_seconds": [
        3,
        8
      ],
      "max_instagram_pages_per_round": 150,
      "wait_after_load_ms": 3500,
      "read": [
        "meta[property='og:description'] -> follower display, e.g. '21K 位粉絲' or '5,881 位粉絲'",
        "bio text from the page body (redact emails/phones before saving)",
        "up to 12 post/reel shortcodes linked on the profile"
      ],
      "latest_post_date": "For each shortcode: decode the first 11 chars with alphabet 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-_' into an integer media id; timestamp_ms = (id >> 23) + 1314220021721. Latest post = max date. No extra page loads.",
      "not_found_markers": [
        "無法顯示",
        "此連結可能故障"
      ],
      "stop_markers": [
        "final URL contains accounts/login",
        "請稍候幾分鐘",
        "HTTP 429"
      ],
      "evidence": "save one screenshot per handle",
      "volume_tested": "78 candidate profiles + 48 post pages = 126 pages, 20 minutes, no login wall, no 429"
    },
    "step_3_judge": {
      "pass_requires_all": [
        "account exists",
        "followers 1,000-50,000 as shown on the public page (never use counts from search results, Manus or other tools)",
        "Taiwan evidence: Taiwan school or city, Taiwan exams (學測/會考/統測/國考/GSAT/CAP), Taiwan university, or 'Taiwan' in bio",
        "current content is study / learning",
        "individual creator (student or teacher-creator); not a countdown page, school unit, publisher, cram school or brand",
        "latest post within 12 months of the check date"
      ],
      "verdicts": {
        "pass": "meets all criteria",
        "partial": "Taiwan evidence weak, topic mixed, or a non-personal but highly relevant community page (e.g. exam countdown pages)",
        "fail-range": "followers outside 1,000-50,000",
        "fail-inactive": "latest post older than 12 months",
        "fail-notfound": "page does not exist",
        "fail-topic": "not study / learning",
        "fail-not-creator": "organization, brand, publisher, cram school"
      },
      "human_review": "A human reads topic, Taiwan evidence and creator type for every pass before outreach."
    },
    "step_4_output": {
      "files": [
        "candidates.csv",
        "ig_public.json (raw, redacted)",
        "screenshots/<handle>.png",
        "judgement.csv"
      ],
      "judgement_columns": [
        "handle",
        "followers",
        "latest_post",
        "taiwan_evidence",
        "topic",
        "verdict",
        "note"
      ],
      "union": "merge pass rows with earlier rounds; record found_by per handle",
      "report": [
        "candidates",
        "passes",
        "pass rate",
        "coverage of known in-range accounts",
        "new handles vs earlier union",
        "minutes for discovery vs verification",
        "cost"
      ]
    }
  },
  "optional_manus_article_candidates": {
    "when": "Want extra candidates from roundup articles, interviews, third-party KOL sites (the weak spot of method_1).",
    "prompt": "Sub-topic candidate-only prompt in the For AI agents section, under Optional Manus add-on: collect only, no filtering, exclude known handles, deliver candidates.csv.",
    "expected": "about 160 credits and 14 minutes per run (EF1, 2026-10-17)",
    "use_output": "Take only the ig_username column; ignore Manus follower counts and verdicts; verify every handle with method_1 step_2.",
    "api": {
      "base_url": "https://api.manus.ai",
      "auth_header": {
        "x-manus-api-key": "YOUR_MANUS_API_KEY"
      },
      "credits": "GET /v2/usage.availableCredits",
      "create": "POST /v2/task.create body {\"message\":{\"content\":\"<PROMPT>\"}}; never retry create",
      "detail": "GET /v2/task.detail?task_id=<ID>",
      "messages": "GET /v2/task.listMessages?task_id=<ID>&order=desc&limit=100 (attachments[].url are temporary; download immediately)",
      "stop": "POST /v2/task.stop body {\"task_id\":\"<ID>\"}",
      "polling": "every 30 s until status == 'stopped' and has_running_background_jobs == false; waiting_for_event_type == 'cascadeJobCall' is normal; any other waiting type -> stop and ask a human",
      "free_account_profile": "lite (forced)"
    }
  },
  "scripts_on_mac_mini": {
    "root": "/Users/shaokanpi/clawd/docs/research/manus-kol-search/scripts",
    "search_harvest.py": "search_harvest.py queries.tsv OUT.json  (queries.tsv: track<TAB>query per line; Startpage + Yahoo TW; 4-8 s; drops an engine on captcha)",
    "n1b_candidates.py": "n1b_candidates.py RUN_DIR  (raw search JSON -> candidates_n1b.csv + unresolved_posts.csv)",
    "ig_post_owner.py": "ig_post_owner.py OUT.json code...  (post shortcode -> owner handle; 3-8 s; stops on block)",
    "ig_public.py": "ig_public.py OUTDIR [--max 150] handle...  (profile followers, bio, latest post date, screenshot; 3-8 s; stops on login wall/429; redacts emails/phones)",
    "spotcheck.py": "spotcheck.py OUTDIR/ig_public.json > spotcheck_raw.csv",
    "union.py": "union.py  (runs/*/judgement.csv -> union_verified.csv with found_by)",
    "run_task.py": "optional Manus runner with budget cap, lock and ledger; never auto-confirms"
  },
  "measured_2026_10_17": {
    "union_verified_handles": 24,
    "A0_manus_baseline": {
      "minutes": 59.8,
      "credits": 132,
      "delivered": 12,
      "passes": 10,
      "known_in_range": "2/3",
      "union_share": "42%"
    },
    "EF1_manus_subtopics": {
      "minutes": 13.7,
      "credits": 157,
      "delivered": 13,
      "passes": 9,
      "known_in_range": "1/3",
      "union_share": "38%"
    },
    "N1_no_manus": {
      "minutes": 34.0,
      "minutes_discovery": 21.4,
      "minutes_verification": 12.6,
      "manus_credits": 0,
      "claude_usd_estimate": "4-6 at API list price",
      "candidates": 78,
      "passes": 20,
      "known_in_range": "1/3",
      "union_share": "83%",
      "only_found_by_N1": 12
    },
    "B1_manus_ig_native": {
      "passes": 0,
      "note": "Instagram in-app search blocked by login wall; do not use"
    },
    "N2_seed_expand_from_bio_mentions": {
      "candidates": 7,
      "passes": 0,
      "note": "bios are truncated; mentions are mostly brands/organizations; not recommended"
    },
    "known_in_range_accounts": [
      "c2h4_study",
      "ninii.study_",
      "study.msh"
    ],
    "rule_change": "Inactive rule (latest post within 12 months) applied to all runs; study_penguin109, yu._.study.0 (EF1) and shannie.studie (B1) changed to fail-inactive."
  },
  "untested": [
    "Running discovery and verification in parallel (estimated ~20 minutes instead of 34).",
    "Whether Startpage / Yahoo Taiwan start blocking headless browsers at higher volume."
  ],
  "acceptance_checks": [
    "candidate handles unique and normalized (lowercase, no @)",
    "every pass verified on the public page in this round: followers in range, latest post within 12 months",
    "Taiwan evidence, topic and creator type recorded per handle",
    "no emails, phone numbers, keys or tokens in any output",
    "page count per round <= 150; stop reason recorded if the run stopped early",
    "report candidates, passes, known-account coverage, new handles, discovery vs verification minutes, cost"
  ],
  "open_items": [
    "Check Meta Business Suite: is Taiwan selectable as creator location or audience location in Creator Marketplace",
    "Influencers.club trial needs a company email",
    "Decide whether countdown pages / cram schools / publishers (non-KOL but relevant audience) become a separate ad list"
  ]
}